Datasets & Benchmarks

Our datasets are released as benchmark resources — each ships with a data card, machine-readable metadata, and a fixed evaluation protocol so that every reported number is reproducible.

Flagship dataset v11 · September 2026 Dataset article in submission

CTSpinoPelvic1K

802 abdominal CT records with the spine, sacrum, hips, femora, per-level ribs and surgical hardware segmented in one coordinate frame. Built from three public collections and annotated for lumbosacral and thoracolumbar transitional anatomy, the levels where conventional pipelines most often fail.

  • Source collectionsCTSpine1K · CTPelvic1K · TCIA CT COLONOG
  • Label scheme67 identifiers: VerSe vertebrae, sacrum, hips, femora, thirteen ribs per side, lumbar ribs, hardware
  • Transitional anatomy33 records with two-reader Castellvi grades · 18 with an L6 · 16 with a lumbar rib
  • SplitsPatient-grouped, LSTV-stratified five-fold
  • DistributionZenodo archive of record (CC BY-NC-SA 4.0) · Hugging Face mirror
Zenodo archive Hugging Face mirror Code and build chain Dataset paper (arXiv)
Reference resource

Lumbar reference atlas, with the spread

The morphometry a surgeon looks up — body height, endplate width, canal dimensions, pedicle width, disc height, bone density — measured across 802 CT records and published with its distribution, in the same axes the standard reference figures use. The classical curves come from cadaveric series of a few dozen specimens and show no spread, so they cannot tell you whether a given patient is unusual. Values downloadable as CSV.

Benchmark

TotalSegmentator Spine Audit

A structured evaluation of a widely used segmentation tool on transitional and variant anatomy — quantifying where automated labels diverge from expert ground truth.

Pipeline

LSTV Detection Ensemble

A multi-model reference pipeline for detecting lumbosacral transitional vertebrae and Castellvi morphology, released to support variant-aware research and replication.

Methodology

LSTV-Aware Training Recipe

An open training methodology — variant-aware oversampling and per-subgroup evaluation — that makes models accountable for rare anatomy rather than averaging it away.

Built on the shoulders of open science. OpenSpineConsortium resources derive from publicly available collections and are redistributed under the terms of their original licenses. Each dataset card documents origin, consent basis, and intended use.

Use the datasets

The resources are free to use under their stated licenses. We only ask that you cite the dataset and report on the standard splits.